Harness adapters: Safe RL for efficient agentic development (huggingface.co)

A new initiative has been launched featuring "harness adapters," which are specialized small LoRA adapters trained with reinforcement learning (RL) to enhance local agentic development efficiently. Each adapter is tailored to a specific model and agent harness—such as pi, opencode, and hermes—designed to reduce unnecessary tool calls while preserving or improving accuracy during evaluations. This development is significant as it streamlines the training process, making it more efficient by allowing various reinforcement learning environments to be incorporated, including tasks like repair, QA, build, and extend, all verified through a set of rigorous metrics. The technical implications are noteworthy: the framework employs "safeagents," which inspect each tool call made by the model to prevent reward hacking. The use of the GRPO algorithm combined with tree attention facilitates efficient training across different models, like the LFM2.5-24B-A2B and LFM3-1.7B. By utilizing these adapters, developers can easily switch models for their specific needs without reinventing the wheel, ultimately making AI development more adaptable and robust across various applications. These innovations not only contribute to safe and efficient reinforcement learning practices but also encourage collaborative advancements within the AI/ML community.